A Surprising Thing: The Application of Machine Learning Ensembles and Signal Theory to Predict Earnings Surprises

Abstract

Nonlinear classification models can predict future earnings surprises with a high accuracy by using pricing and earnings input data. Surprises of 15% or more can be predicted with 71% accuracy. These predictions can be used to form profitable trading strategies. Additional variables have been created using signal-processing and handcrafted feature-engineering methods. Some of these variables have in the past been known to be related to analyst bias. The machine learning model in effect corrects for analyst mistakes and biases by incorporating these variables into a nonlinear prediction model to predict future earnings surprises.

Keywords: Machine Learning, Earnings Surprise, Event-driven, Trading Strategy, Prediction

JEL Classification: C32, C38, C45, G14